A16 Bionic

[exoticspice1]

Very little GPU improvement in A16. No wonder Apple did say anything new about the GPU.

 

[Jimmyjames]

View attachment 17649
Very little GPU improvement in A16. No wonder Apple did say anything new about the GPU.

A16 Bionic

A dedicated thread to talk about Apple's latest chip :) Seems like this leaked Geekbench result (https://browser.geekbench.com/v5/cpu/17141095) is more in line with expectations: 1887 in single core, 5455 in multicore (+9.2%, +13.4%). Interestingly, the biggest improvement in multicore is in...

talkedabout.com

This has just been discussed. GB has issues with ramping up too slowly on ASi. Antutu shows a much healthier improvement. I have no idea how reliable or reputable Antutu is however!

 

[exoticspice1]

A16 Bionic

A dedicated thread to talk about Apple's latest chip :) Seems like this leaked Geekbench result (https://browser.geekbench.com/v5/cpu/17141095) is more in line with expectations: 1887 in single core, 5455 in multicore (+9.2%, +13.4%). Interestingly, the biggest improvement in multicore is in...

talkedabout.com

This has just been discussed. GB has issues with ramping up too slowly on ASi. Antutu shows a much healthier improvement. I have no idea how reliable or reputable Antutu is however!

That only applies to Mx Max and Ultra chipsets.

Geekbench is fine to use for the A series and Mx base series. Previous A series chips in the GPU test have shown huge improvements. For example the A14 iPhone 12 Pro scored around 9300 and A15 on iPhone 13 Pro scored around 14500. So ultra mobile chipsets Geekbench is fine to use.

 

[Andropov]

Hence I'd expect that the same frame on a A16 Bionic with +50% memory bandwidth to be around (gross estimate) ~9% faster due to memory bandwidth improvements alone.

Very little GPU improvement in A16. No wonder Apple did say anything new about the GPU.

That's a 14509 -> 15807 = +8.9% in GPU performance... almost exactly what I estimated from bandwidth improvements alone. Hmm.

 

[Jimmyjames]

That only applies to Mx Max and Ultra chipsets.

Geekbench is fine to use for the A series and Mx base series. Previous A series chips in the GPU test have shown huge improvements. For example the A14 iPhone 12 Pro scored around 9300 and A15 on iPhone 13 Pro scored around 14500. So ultra mobile chipsets Geekbench is fine to use.

Hmmmm. We don’t really know that it does scale well on the A series. It’s true that previous generations showed a bigger leap, but the issue was ramping up regardless of chip size. It’s quite possible that it underestimates performance on all Apple Silicon chips, and that some modifications would yield hiher scores all round. Hence the better performance on other benchmarks like gfxbench and Antutu.

 

[exoticspice1]

That's a 14509 -> 15807 = +8.9% in GPU performance... almost exactly what I estimated from bandwidth improvements alone. Hmm.

Thats because the GPU is the same. I don't see M3 being based off A16 GPU and M2 already has LPDDR5.

M2 GPU was based of A15 but M3 GPU maybe based of A17.

 

[theorist9]

Are single-precision (FP 32) TFLOPS a good meansure of general GPU compute performance? If so, and based on this, I'd say get a Mac if your primary concern is GPU performance/watt, or go PC-NVIDIA if your primary concern is GPU performance/$*:

[I don't fall into either category; for me, the OS and my attendant user efficiency and experience are paramount, so I use a Mac.]

TFLOPS, SINGLE-PRECISION (FP 32)
M1: 2.6
M2: 3.6
M1 MAX: 10.4
M2 MAX: 14 (?) (EXTRAPOLATING FROM M2/M1 x M1 MAX)
4050: 14 (?) (entry-level, ~$250?)
M1 ULTRA: 21
M2 ULTRA: 29 (?) (EXTRAPOLATING FROM M2/M1 x M1 ULTRA)
3080: 30
4060: 31 (?) (entry-level, ~$330?)
3080 TI: 34
3090: 36
3090 TI: 40
4070: 43 (?) (mid-level, ~$550?)
4080: 49 (?)
M2 EXTREME: 58 (?) (EXTRAPOLATING FROM 2 x M2 ULTRA)
4090: 80 (?)
4090 TI: 89 (??) (EXTRAPOLATING FROM 4090 x 3090 TI/3090)
M2 2X EXTREME: 116 (?)

*Of course there is also the energy cost. Don't know what the actual efficiencies are, but if a 500 W NVIDIA GPU is twice as fast as a 100 W AS GPU, and thus you'd need to run the 100 W GPU 5 hrs/day vs. 2.5 hours/day for the 500 W GPU, then the 500 W GPU would cost an extra $27/year for every 10 cents/kwh that your electricity costs (if both were run 365 days/year).

 

[Colstan]

Concerning Geekbench, it's almost like relying upon a single benchmark to determine an entire platform's capabilities is a fool's errand. I realize that it's the best we have right now to get an idea of A16's performance, but it's woefully inadequate in determining how much different it is in regard to its predecessor and competitors. It would be nice if we could run a two minute series of benchmarks and know everything there is to know about a device, but it doesn't work that way.

If you've been following the history of the industry for long enough, you'd know that individual benchmarks have issues, even going as far to cheat in order to favor one company over another, which has happened repeatedly in this business. I'm not accusing Geekbench of any shady shenanigans, they appear to produce the most accurate results that they can, but a single benchmark is never going to tell the whole story. It's difficult to find truly cross-platform, widely available results, but this needs to be viewed with caution until more data is available. Plus, we don't know exactly what features Apple has prioritized for the iPhone, and how this will impact the M-series going forward.

[I don't fall into either category; for me, the OS and my attendant user efficiency and experience are paramount, so I use a Mac.]

I completely agree with @theorist9. I pay attention to what the PC crowd are doing, because I am interested in technology. However, there's absolutely nothing that Microsoft, Intel, AMD, or Nvidia can release that's going to persuade me to switch from the Mac. In fact, with the way PCs are getting hotter, noisier, and guzzling more energy, they're heading in the opposite direction. It doesn't help that the operating system and ISA are covered with barnacles accumulated through decades of mismanagement, with the inability to cut out cruft and dead weight.

So, even if Apple doesn't always win the raw performance race every time, I'm okay with that. I'm not using a graphics card or CPU in isolation. I'm using a computer to accomplish real tasks. I appreciate that that Apple controls the entire stack, from the microcontrollers that drive the I/O ports, to the SoC, to the operating system and primary applications. It's the vertical integration strategy that Steve Jobs could have only dreamed of for the Mac. That's why I'm an Apple customer, not because it has the highest TFLOPS in a synthetic benchmark, but because I consider the user experience to be the absolute best.

 

[mr_roboto]

Geekbench isn't actually a single benchmark. It's a self-contained suite of many benchmarks. It generally appears to have been inspired by SPEC, and has been shown to correlate very well with SPEC.

 

[Cmaier]

Geekbench isn't actually a single benchmark. It's a self-contained suite of many benchmarks. It generally appears to have been inspired by SPEC, and has been shown to correlate very well with SPEC.

it’s a single benchmark in the sense that it collects the results of multiple workloads and produces a single score which is what most people look at.

even chip designers, who think in terms of SPEC, do not rely just on SPEC. You want a whole array of benchmarks from different sources with different philosophies, different workloads, different runtimes, etc.

 

[tomO2013]

I’m sitting on a bus with a 12 hour bus journey ahead of me. The person next to me has run out of juice twice and currently is using a large power bank to get juice to her Lenovo. I’m still working away on my macbook!

It’s easy to get distracted with who has got the fastest, who has got the biggest numbers in an individual benchmark etc…
However day to day living with the product (for my needs) performance per watt trumps all else.
Perhaps it’s a little smug of me to say this, but I have a smile on my face that I know a secret that the lady next to me on the bus doesn’t. My macbook pro is unplugged, with juice to spare and still on it’s first charge while plowing through XCode tasks, multiple docker containers.
Any focus on efficiency and news of A16 is very exciting to me. Geekbench is interesting when you consider it from a traditional cpu vendor perspective. Apple however is NOT a traditional CPU vendor! They don’t sell CPU’s. They have no real interest beyond the peppering some bench’s as part of a wholistic product launch like a macbook. Intel, AMD on the other hand , cell CPU’s as the product.
You really have to question generic cross platform compiled benchmarks when viewed through the lens that Apple silicon is sold as part of a product platform. It’s never intended to be used to run various versions of Windows, different flavours of Linux (Redhat, CentOS, Mint, Ubuntu, etc…). It’s designed to run on a controlled ecosystem - either iOS, iPad OS, Mac OS. The co-processors and accelerators that apple includes on die, are designed for the most part to be accessed through Apples provided development tooling workflow and be as transparent as possible to the developer who uses their Core Frameworks etc… Geekbench will never be able to show the benefits of this approach e.g. 8k smooth playback in FCP or DaVinci Resolve because to code such a benchmark would mean ignoring the co-processors that Apple provides and instead doing everything via generic CPU/GPU calls. Thats not to say Geekbench is useless, it’s more to say that it needs to be considered a datapoint and not representative of the platform performance holistically which I feel many youtube reviewers (and purveyors of random crab cake chess and hungry hungry hippo benchmarks over at the other place) tend to view it.
In other words, in my humble opinion, Apple Silicon a design philosophy that is not intended to target cross platform generic code and instead is built around the philosophy of optimization where Apple controls the full widget.
The A-series are designed to be optimized for efficiency (perf/watt) and an optimized application workflow in the real world.
If we are not seeing much beyond 9% improvement in GPU with A16 , we can likely be reasonably sure that the GPU itself is sipping less power from a perf/watt perspective relative to A15. That’s great news!
As an aside, NVidias latest GPU 4090 is rumored to have a 660W TDP for the graphics card alone. An individual part.
Apple’s approach with TBDR is a smarter approach to GPU efficiency and performance (in a world that we live in that is more energy conscious)
Just my 0.02 from my perspective.
Hope y’all having a great day.

 

[Cmaier]

The thing with benchmarks is more or less the thing with standardized school testing. You get what you test for. If you test and quantize A, B, and C and publish a ranking based on that test, you end up getting everyone competing in that ranking. But to a particular user, D, E, and F may be what is important.

At AMD we had thousands of “traces” that we looked at (for compatibility and performance testing). You can’t optimize for everything, but you do want to have a very broad spectrum of things that you at least look at. Some of the standard benchmarks back in the day went to things like processing a SPICE deck (a tool used by electrical engineers), which is a pretty specialized use case that had an outsized impact on how CPU vendors competed with each other back in the day, for example.

 

[mr_roboto]

it’s a single benchmark in the sense that it collects the results of multiple workloads and produces a single score which is what most people look at.

even chip designers, who think in terms of SPEC, do not rely just on SPEC. You want a whole array of benchmarks from different sources with different philosophies, different workloads, different runtimes, etc.

Sure, was just saying that it's got a variety of loads under the hood.

And you're absolutely right that SPEC isn't enough - from a 1990s UNIX workstation perspective, SPEC is reasonably broad, but from a general purpose desktop OS perspective it's not. If nothing else, most of SPEC is C and a little bit of Fortran, and we're in the middle of a slow shift away from C-family languages. (Which is a somewhat faster shift on Apple's own platform, given how much they're pushing Swift.)

 

[Yoused]

Are single-precision (FP 32) TFLOPS a good meansure of general GPU compute performance?

Kind of not a simple question. For instance, in the late '90s, Apple had the fastest (MHz) computer on the consumer market. It ran Mac OS 8+, which, IME, would crash once or twice a day, forcing you to reboot – in today's environment, classic Mac OS would be dreadfully compromised and simply untenable. Windows 9x seemed like it should be safer, but the sloppy coding methods they used made it actually worse, and when a Windows PC crashed, more often than not, the typical user would end up doing a nuke-and-pave on the HD, which was almost never necessary on a classic Mac.

What I am rambling on about is that software/system design is at least as important as mips/spec/gb5/cb23/etc. I believe that Apple will pursue a strategy where GPU TFLOPs are offset by a more elegant approach. I see where Tom's says the A16 neural engine performanre is 40% improved – which adds, what? 40% is quite a lot. My notion is that they will leverage ML to teach the machine to work smarter, not waste effort on large, homogenous swaths of image but focus on rendering the complex parts. If you can do effectively the same job for less effort, that saves juice and maybe time as well.

 

[exoticspice1]

Gfxbench shows no improvement. So A16 and A15 have the same GPU.

 

[Cmaier]

Gfxbench shows no improvement. So A16 and A15 have the same GPU.

Perhaps. But if a benchmark shows no improvement it’s a questionable benchmark - the increased bandwidth should cause an increase in performance.

 

[Andropov]

Gfxbench shows no improvement. So A16 and A15 have the same GPU.

Where have you found a Gfxbench benchmark of the A16?

 

[exoticspice1]

Perhaps. But if a benchmark shows no improvement it’s a questionable benchmark - the increased bandwidth should cause an increase in performance.

But that same benchmark had showed increased pref before you know when Apple actually improved the GPU.
Like from A14 > A15.

For the first time Apple's GPU did not get new cores.

 

[exoticspice1]

Where have you found a Gfxbench benchmark of the A16?

source:

 

[Jimmyjames]

Not what I’m seeing. Not sure why the gfxbench listed excluded 4K.